Math in latex format in markdown files

[AnirudhG07]

Is there a way to generate mathematics in proper latex format extracted from pdf's or other types of files? I am open to using LLM's with it, but I want markitdown in the loop instead of just raw LLM prompt with attached pdf.

[ltianyi992]

markitdown does not currently extract math as LaTeX — this is a known gap when working with scientific PDFs. Here's a breakdown of what works and what the alternatives are:

Why markitdown struggles with math:
markitdown uses pdfminer for PDF text extraction, which reads character streams. Mathematical symbols get extracted as Unicode characters (∑, ∫, ∂, etc.) rather than as structured LaTeX expressions. Equations in image form (scanned or vector-embedded) are skipped entirely.

Option 1 — LLM integration (markitdown stays in the loop)
markitdown supports passing an llm_client for image captioning. You can use this for per-page images to ask the LLM to render math as LaTeX:

from openai import OpenAI
from markitdown import MarkItDown

client = OpenAI()
md = MarkItDown(
    llm_client=client,
    llm_model="gpt-4o",
)
# Convert PDF — images (including math figures) get described by the LLM
result = md.convert("paper.pdf")
print(result.text_content)

This works best for figure-based equations. For inline text math, the PDF text layer is still used as-is.

Option 2 — Nougat (Meta, best for scientific PDFs)
Nougat is purpose-built for academic PDFs and outputs mmd (Myst Markdown) with proper LaTeX math fences. It stays fully local:

pip install nougat-ocr
nougat paper.pdf -o output/
# output/paper.mmd contains $$...$$  and $...$ LaTeX

You can then pass the .mmd output to markitdown (it handles Markdown input) or use it directly.

Option 3 — MathPix API (cloud, highest accuracy)
MathPix is a paid API specialized in math OCR — it converts PDFs/images to LaTeX with very high accuracy for complex notation.

Option 4 — pymupdf4llm with math hints
pymupdf4llm extracts text with better structure than pdfminer and preserves some symbol context that an LLM can then interpret as LaTeX.

Recommended pipeline for keeping markitdown in the loop:

1. Use Nougat or MathPix to get a .mmd/LaTeX intermediate
2. Pass that to MarkItDown().convert_string(text) or feed it directly to your LLM

This gives you LaTeX math while still using markitdown's chunking and conversion infrastructure.